Polls and the Vote in Britain - Christopher Wlezien
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
bs_bs_banner
P O L I T I C A L S T U D I E S : 2 0 1 3 VO L 6 1 ( S 1 ) , 6 6 – 9 1
doi: 10.1111/1467-9248.12008
Polls and the Vote in Britain
Christopher Wlezien Will Jennings Stephen Fisher
Temple University University of Southampton University of Oxford
Robert Ford Mark Pickup
University of Manchester Simon Fraser University
Little is known about the evolution of electoral sentiment over British election cycles. How does party support
converge on the eventual election outcome? Do preferences evolve in a patterned and understandable way? What
role does the official election campaign period play? In this article, we begin to address these issues. We outline an
empirical analysis relating poll results over the course of the election cycle and the final vote for the three main
political parties. Then we examine the relationship relying on vote intention polls for the seventeen British general
elections between 1950 and 2010. Predictably, polls become increasingly informative about the vote over the
election cycle. More surprisingly, early polls contain substantial information about the final outcome, much more
than we see in presidential and congressional elections in the US. The final outcome in Britain comes into focus
over the long campaign and is to a large extent in place well before the official election campaign begins. The
findings are understandable, we think, but raise other questions, which we begin to consider in a concluding
section.
Keywords: election; forecasting; campaigns; parliamentary systems; presidential systems
Previous research shows that opinion polls are informative about upcoming general election
outcomes in Britain. This is especially true on election eve, though even then their
performance has been uneven – with particularly poor performances in 1992 and 2010
(Jowell et al., 1993; Market Research Society, 1994; Pickup et al., 2011). Naturally we expect
polls closer to the election to correspond more closely to the eventual outcome than those
further away, but the particular course that polls follow as an election nears can provide us
with important insights about how this happens. We want to know how much more
informative polls become over the course of the election cycle.
It may be that polls have little information content before the formal election campaign
begins, but rapidly become informative during the campaign as the public focuses on the
election and views of the parties and candidates become more settled. Alternatively, it may
be that polls have high information content from much earlier in the political cycle,
suggesting that public opinion takes shape long before voters and parties formally focus on
the election itself. Further, the pattern could vary by party.
The extent to which preferences coalesce before the official campaign begins has direct
implications for the influence of election campaigns. It sheds light on the potential to
inform voters and shape preferences and the timing within the electoral cycle when this
may occur (Andersen et al., 2005;Arceneaux, 2005; Finkel, 1993; Hillygus and Shields, 2008;
Matthews and Johnston, 2010; Stevenson and Vavreck, 2000; for a review of the literature,
see Hillygus, 2010). Relatedly, it reveals limits on the voter response to major events and
© 2013 The Authors. Political Studies © 2013 Political Studies AssociationP O L L S A N D T H E VO T E I N B R I T A I N 67
policy announcements and the extent to which the effects of these remain to impact the
final electoral vote (Holbrook, 1996; Shaw, 1999; Erikson and Wlezien, 2012).
The institutions and party system of a country are likely to structure how electoral
preferences evolve, by influencing the stability of the choices faced by voters and the level
of information they have about choices at different points in the cycle. Examining the
predictive power of polling in different countries can thus provide us with important
insights about how the evolution of political choice differs in different institutional
frameworks and political traditions. In the next section, we consider ways in which the
evolution of preferences could be influenced by such factors, and propose the key hypo-
theses that we examine with our data.
Theory and Hypotheses
Political institutions can structure the evolution of public opinion in several ways. First,
in Westminster systems, we expect early polls to be informative about the election
outcome. This actually should be true in all parliamentary systems, as national parties
predominate – since voters’ party preferences tend to be fairly stable over time, early polls
should be informative. Now in Westminster systems, the government almost always is
controlled by a single party, and voters decide to continue with that party or change
course to go with the alternative.This, and the fact that governments in Britain, like most
parliamentary systems, can call an election at any point in time, encourages voters to
update their evaluations of parties throughout the electoral cycle. That is, the possibility
of a new election is always present, at least to some degree. As a result, there is likely
to be a greater focus among politicians and the media on the electoral horse race,
and British voters are more likely to keep a ‘running tally’ of party performance. If
so, polls would reveal ‘enlightened preferences’ (Gelman and King, 1993) early in the
cycle.
H1. Polling in Britain should be informative about election outcomes from early in the
election cycle.
Second, the informativeness of the polling for a particular party may be influenced by the
structural position of the different parties in the political system. There are a number of
factors that might contribute to party differences in how informative polls are at particular
stages in the electoral cycle.When certain parties dominate a system as the only ones likely
to feature in government, attention naturally tends to gravitate to these parties. Third and
minor parties tend to be marginalised outside election time, receiving less media coverage
and possessing a limited ability to set the agenda. In the British context, we might expect
that polls a long way from the election should be more informative for the dominant
Labour and Conservative Parties, and much less informative for the large, but structurally
disadvantaged, third party – the Liberals (including their successors the Alliance and Liberal
Democrats).This disadvantage may be much reduced during the formal campaign period,
when media organisations are obliged to provide equal coverage of all three large parties
(e.g. Fieldhouse and Cutts, 2009;Whiteley and Seyd, 2003). So there is reason to expect that
the predictive power of Liberal polls will improve rapidly during formal campaign periods,
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)68 C H R I S T O P H E R W L E Z I E N ET AL.
as voters acquire new information about the previously under-reported party and adjust
preferences accordingly.1
There are other reasons, too, for expecting party differences in the evolution of electoral
preferences more generally. One possibility is that tactical voting might contribute to late
movements of support.There could, alternatively, be greater churn between the supporters
of certain parties than others (for example, such as the competition between Labour and the
Liberals for left-of-centre voters since the days of the SDP–Liberal alliance), leading
electoral preferences to come into focus quite late for those parties. For many reasons, then,
we expect party differences in the pattern of change in informativeness of polls over the
election cycle.
H2. There should be differences among political parties in the degree to which polling is
informative about election outcomes, over the course of the election cycle.
Our third expectation is that polls are likely to be informative from much earlier in the
electoral cycle in parliamentary systems such as Britain than in presidential systems such as
the US.There are two main reasons for this. First, as we have discussed, British elections are
substantially centred on national parties. In presidential systems, by contrast, candidates
matter more, though especially in the presidential ballot. In the US, for instance, at least one
of the presidential candidates is not decided until late in the election cycle. Early polling in
parliamentary systems therefore should be more informative than presidential polling,
although this differential should rapidly dissipate once the identities of the presidential
candidates are known. Second, even to the extent that British party leaders matter (Stewart
and Clarke, 1992), they typically are known well in advance of the election. This allows
voters to take them into account early on in their electoral preferences. In the US, by
contrast, the presidential candidates are the de facto party leaders and are not fully known
until months before the election. Voters thus cannot take into account the presidential
candidates in their congressional election preferences until very late in the election cycle.
Early polling in Britain is expected to be more informative than both presidential and
congressional polling in the US.
H3. Polling in Britain should be more informative early in the cycle than executive or
legislative polling in the US, and the differential should be particularly large before the
identities of presidential candidates become known.
In this article we test these hypotheses by examining vote intention polls for the seventeen
British general elections between 1950 and 2010.We assess the movement in aggregate poll
results over each election cycle. We then analyse the relationship between poll results at
different points in time and the final election day vote. Finally, results from the UK are
compared with results for presidential and congressional elections in the US. Our results
provide supporting evidence for each of our hypotheses.
The Polls, 1945–2010
We have compiled what we believe is the most complete data set of available national polls
of the vote division that include the Conservative, Labour and Liberal Parties for the
seventeen elections between 1950 and 2010.2 Specifically, we have 3,293 opinion polls from
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)P O L L S A N D T H E VO T E I N B R I T A I N 69 eighteen polling ‘houses’ – that is, companies. (Details are reported in the Appendix.) In the polls, respondents were asked about how they would vote ‘if the election were held tomorrow’ with minor variations in question wording. In analysing these polls, we ignore these differences in wording. For our analyses, we use the published ‘headline figures’ released by polling companies as their current estimate of vote intention in Great Britain. There are two important points to note about these data. First, we are compelled to work with headline vote intention figures, which do not reflect consistent sampling or weighting strategies by different polling organisations or even by the same organisation over time. Going back in time, most opinion polls in Britain up until the 1990s were conducted using face-to-face quota or random samples, with telephone surveys first used at the 1979 election (Worcester, 1980). Only since the 1990s have pollsters made regular use of telephone polling, with internet panels becoming increasingly widespread since the 2000s, leading to increasing variation in the weighting and filtering strategies used in polling. Today, British polling organisations employ different strategies for weighting their data by likelihood to vote, partisanship and other things, all of which differ between organisations and may change over time. The headline figures may thus reflect different sampling universes. Practices also evolve over time – past vote weights, for example, were mainly introduced after the 1992 election, when the polls performed particularly poorly (Jowell et al., 1993). While ideally we would like to work with consistent data, assembling a time series reflecting consistent weighting and sampling practices would be impossible. We therefore use the ‘headline figures’ vote intentions as the most consistent attainable time series of poll data, one that reflects the survey houses’ best estimate of vote preferences at the time of the survey.Where a survey house changes its sampling or weighting strategies, our poll data will therefore capture this change.3 Second, especially in recent years, survey organisations often report results for overlap- ping polling periods.This is quite understandable and is as we would expect, for example, where a survey house operates a tracking poll and reports three-day moving averages. Respondents interviewed today would be included in the published poll results for the following two days. Clearly, we do not want to count the same respondents on multiple days, and it is very easy to remove this overlap. For the hypothetical survey house operating a tracking poll and reporting three-day moving averages, we would only use poll results for every third day. For the period 1950–2010, this leaves 3,033 polls of vote intention.Table 1 summarises the data by election cycle. For instance, in the uppermost row, there were 32 polls conducted after the election of 1945 and before the election of 1950. In the table we can see that around two-thirds of the polls were conducted during the last six election cycles – 1987–2010. During this period, the number of polls averaged 344 per election; before- hand, we only have 87 polls per election on average. The number of polls varies from election to election for other reasons too, and the length of the inter-election period is one of them, that is, when there is less time between elections, there are fewer polls. For example, there were only 36 polls conducted prior to the second 1974 election, which occurred only eight months after the first election of that year, and 48 polls during the seventeen months between the 1964 and 1966 elections. © 2013 The Authors. Political Studies © 2013 Political Studies Association POLITICAL STUDIES: 2013, 61(S1)
70 C H R I S T O P H E R W L E Z I E N ET AL.
Table 1: A Summary of the British Poll Data, 1950–2010
Election year Election called Election day Number of pollsa Number of daysb
1950 11 January 23 February 32 32
1951 19 September 25 October 18 18
1955 15 April 26 May 40 40
1959 8 September 8 October 59 59
1964 15 September 15 October 134 129
1966 28 February 31 March 48 48
1970 18 May 18 June 130 123
1974 (February) 7 February 28 February 89 81
1974 (October) 18 September 10 October 36 29
1979 29 March 3 May 165 158
1983 9 May 9 June 210 180
1987 11 May 11 June 284 254
1992 11 March 9 April 482 406
1997 17 March 1 May 300 266
2001 8 May 7 June 210 190
2005 5 April 5 May 229 179
2010 6 April 6 May 559 448
Total 3,033 2,640
Notes: a The number of pre-election polls conducted since the previous election. b The number of days since the previous election
for which poll readings are available, based on an aggregation of polls by the mid-date of the reported polling period.
Since most polls are conducted over multiple days, we ‘date’ each poll by the middle day
of the period the survey is in the field.4 Using this method, the 3,033 polls allow readings
(sometimes multiple) for 2,640 separate days from 1945 to 2010 (see Table 1). For days
when more than one poll result is recorded, we pool the results together into one poll of polls
by taking the (unweighted) mean. Since the 1987 election, we have readings for 291 days
per election on average. During the official election campaign in these years, we have a
virtual day-to-day monitoring of preferences, especially since 1992.5
Figure 1 presents the daily poll of polls for the full election cycle in each of the seventeen
elections between 1950 and 2010.6 The figure depicts each of the main three parties’ share
of vote intention, ignoring all other parties, aggregated by the mid-date of the reported
polling period. We notice a number of things in the figure.
First, we see that the length of the polling period varies quite a lot from election to
election. This to a large extent reflects the length of the inter-election period, as noted
above. Second, the variance of electoral preferences differs quite a lot from election to
election. During the 1983 cycle the polls shifted dramatically for all three parties. In the
1970 and 1997 cycles, we also see substantial churning in support for the Conservatives and
Labour. In the early elections (1950, 1951 and 1955), conversely, preferences shifted in a
narrow range. Third, just as the variance of the poll time series differs from election to
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)Figure 1: Vote Intention during the Electoral Cycle, 1945–2010
1950 1951 1955 1959 1964
60 60 60 60 60
50 50 50 50 50
40 40 40 40 40
30 30 30 30 30
20 20 20 20 20
10 10 10 10 10
POLITICAL STUDIES: 2013, 61(S1)
0 0 0 0 0
2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0
Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon
1966 1970 1974 (Feb) 1974 (Oct) 1979
60 60 60 60 60
50 50 50 50 50
40 40 40 40 40
30 30 30 30 30
© 2013 The Authors. Political Studies © 2013 Political Studies Association
20 20 20 20 20
10 10 10 10 10
P O L L S A N D T H E VO T E I N B R I T A I N
0 0 0 0 0
2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0
Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon
1983 1987 1992 1997 2001
60 60 60 60 60
50 50 50 50 50
40 40 40 40 40
30 30 30 30 30
20 20 20 20 20
10 10 10 10 10
0 0 0 0 0
2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0 2000 1500 1000 500 0
Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon
2005 2010
60 60
50 50
40 40
30 30 ConservaƟve Labour Liberal
20 20
10 10
0 0
2000 1500 1000 500 0 2000 1500 1000 500 0
Days UnƟl ElecƟon Days UnƟl ElecƟon
7172 C H R I S T O P H E R W L E Z I E N ET AL.
election, the within-cycle variance of polls appears to differ from one time period to
another over the campaign.That is, the volatility of the polls drops noticeably as the election
cycle evolves.
Figure 2 zooms in on the official campaign period, and displays the daily poll of polls for
the final 30 days prior to election day. Here we can see that preferences shift within a much
narrower range. Perhaps more importantly, the party leading at the beginning of the
campaign never loses the lead in the polls, though in some years (1959, 1979, 1992 and
2005) it gets very close and in one year (1992) the leader in the polls actually lost the
election. Indeed, there appears to be a general tendency for poll leads to shrink during the
official campaign. At the same time, perhaps because of the increased media exposure
mentioned above and because the coverage tends to be positive (Norris et al., 1999), the
Liberals almost always gain – specifically, their share increases in thirteen of the fifteen
election cycles for which we have data, by 2.5 percentage points on average.7
Tables 2 and 3 present the daily variance in the polls during the seventeen election cycles.
Table 2 shows the variance over the full election cycle and Table 3 the variance over the
final 30 days. The numbers confirm what we observed in Figures 1 and 2. Poll variance
clearly differs a lot across elections, being more pronounced in some years than others. It
also differs across the election cycle, being more pronounced earlier in the cycle than later.
On average, the variance of poll preferences during the official campaign is only about
one-quarter of the variance for observations over the full election cycle. There are some
interesting differences across the political parties, however. While the variance of polls
declines markedly for all parties, the drop-off is much less pronounced for the Liberals.That
is, during the official campaign, polls of Liberal support vary at least as much as polls of
Labour and Conservative support, while otherwise the Liberal poll variation is lower than
that of the two main parties.This is consistent with our hypothesis that the campaign period
is of special importance to the Liberal vote.The increased attention media organisations are
obliged to give to the party, its candidates, local campaigning and its leader during the
campaign seemingly help to offset the structural disadvantages the party suffers outside
campaign periods.
The numbers in Table 3 suggest that the patterns have changed somewhat over time, and
that poll preferences during recent cycles are more volatile than they were in the 1950s and
1960s. The mean variance (calculated across the three parties) prior to the 1979 election
cycle is 2.9 percentage points; since that time, it is 5.2 percentage points.8 This may be
something of a surprise given the substantial increase in the number of polls during the later
period, which has increased the number of respondents and should dampen the variance
due to sampling error.The increased volatility of the polls may reflect the greater diversity
of polling houses in recent election years, which may have introduced greater house effects
and poll variance (see Note 6).There are other possible explanations, including the decline
in the strength of British voters’ party attachments over time (Clarke et al., 2004; Dalton and
Wattenberg, 2000; Schmitt, 2002).
The Methods
Polling tells us how people would vote if the election were held ‘tomorrow’, not how they
will vote when the election actually occurs. Between the date of the poll and election day
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)Figure 2: Vote Intention during the Official Campaign: The Final 30 Days, 1945–2010
1955 1959 1964 1966
60 60 60 60
50 50 50 50
40 40 40 40
30 30 30 30
20 20 20 20
10 10 10 10
0 0 0 0
POLITICAL STUDIES: 2013, 61(S1)
30 25 20 15 10 5 0 30 25 20 15 10 5 0 30 25 20 15 10 5 0 30 25 20 15 10 5 0
Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon
1970 1974 (Feb) 1974 (Oct) 1979
60 60 60 60
50 50 50 50
40 40 40 40
30 30 30 30
20 20 20 20
10 10 10 10
0 0 0 0
© 2013 The Authors. Political Studies © 2013 Political Studies Association
30 25 20 15 10 5 0 30 25 20 15 10 5 0 30 25 20 15 10 5 0 30 25 20 15 10 5 0
P O L L S A N D T H E VO T E I N B R I T A I N
Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon
1983 1987 1992 1997
60 60 60 60
50 50 50 50
40 40 40 40
30 30 30 30
20 20 20 20
10 10 10 10
0 0 0 0
30 25 20 15 10 5 0 30 25 20 15 10 5 0 30 25 20 15 10 5 0 30 25 20 15 10 5 0
Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon
2001 2005 2010
60 60 60
50 50 50
40 40 40
30 30 30
20 20 20
10 10 10
0 0 0
30 25 20 15 10 5 0 30 25 20 15 10 5 0 30 25 20 15 10 5 0
Days UnƟl ElecƟon Days UnƟl ElecƟon Days UnƟl ElecƟon
ConservaƟve Labour Liberal
7374 C H R I S T O P H E R W L E Z I E N ET AL.
Table 2: Variance of Daily Reports of Vote Intention,
1950–2010
Election year Conservative Labour Liberal
1950 10.0 6.1 2.6
1951 13.7 13.3 1.6
1955 7.6 2.2 4.4
1959 14.0 6.9 10.4
1964 19.8 15.8 20.3
1966 7.3 11.4 3.1
1970 28.6 46.8 4.4
1974 (February) 14.9 20.2 35.8
1974 (October) 3.4 6.5 3.7
1979 17.9 17.0 7.4
1983 39.2 55.3 71.1
1987 21.5 11.5 20.3
1992 22.2 24.4 14.2
1997 25.1 34.1 14.6
2001 12.2 28.1 5.0
2005 9.4 26.7 7.5
2010 14.2 21.9 10.0
Mean 16.5 20.5 13.9
Median 14.2 17.0 7.5
much can happen to change voters’ minds, including the short and intense formal campaign
period of approximately a month between the announcement of an election and the date
on which voters cast their ballots. Some of the change in electoral preferences is impossible
to anticipate, as it is due to unique shocks and events.
Some of the change, however, may be driven by systematic and predictable effects. It may
be, for instance, that pre-election poll leads tend to fade and gravitate towards some
equilibrium preference.9 Such ‘regression to the mean’ is observed in a variety of social and
physical phenomena (Barnett et al., 2005). Measurements taken at one point in time that are
unusually large or small revert to something closer to the average by the time of the next
measurement.This can be due both to measurement error and to underlying processes that
produce mean reversion. So it could be with public opinion and elections. If a party is doing
unusually well in the polls, it may be expected, based on the experience of past election
cycles, that the party will lose support leading up to the election, and for parties that
are doing unusually poorly in the polls to gain support.The election day outcome would
be the equilibrium plus the effects of late campaign events that do not fully decay before
the vote.
It may be that events have more permanent effects. In the extreme, all effects would last to
influence the outcome, and the election day result would simply sum the effects of all events
over the election cycle.10 If true, polls would become increasingly informative as the election
cycle unfolds and preferences converge on the final result. A more reasonable expectation
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)P O L L S A N D T H E VO T E I N B R I T A I N 75
Table 3: Daily Variance of Vote Intention during the Last 30
Days of the Cycle, 1950–2010
Election year Conservative Labour Liberal
1950 – – –
1951 – – –
1955 1.5 0.1 0.6
1959 3.3 2.9 1.9
1964 0.8 1.4 0.1
1966 0.8 1.3 0.3
1970 4.1 1.5 1.6
1974 (February) 6.5 7.2 16.0
1974 (October) 3.8 2.6 2.2
1979 11.8 4.9 4.9
1983 4.0 10.6 13.4
1987 0.9 2.3 2.6
1992 2.4 2.1 3.7
1997 3.6 7.5 5.4
2001 2.6 9.4 4.0
2005 2.1 1.7 1.1
2010 4.6 2.9 15.7
Mean 3.5 3.9 4.9
Median 3.3 2.6 2.6
may be that not all effects persist, but rather some last and others decay. Here, there would be
a moving equilibrium above or below which preferences may shift in the short run; that is,
there would be regression to a moving mean.The election day outcome would be the final
value of the moving equilibrium plus the effects of late campaign events that do not fully
decay before the vote. Under these circumstances, the polls would increasingly converge on
the final result, especially towards the very end of the election cycle.11
In theory, the time-series type could be inferred from standard statistical diagnostics on
time series of polls.That is, we could make the determination by estimating the following
equation for each party in each election cycle:
VOTE INTENTIONt = α + β VOTE INTENTIONt −1 + εt , (1)
where vote intention is the party’s share in the polls at each point in time t and e is a series
of independent shocks drawn from a normal distribution.12 The parameter b reveals the
main time-series characteristic of the series – whether observed changes to preferences
persist or decay and, if the latter, the rate of decay. Two problems complicate such an
analysis. First, as we noted earlier, a non-trivial portion of the observed variance in party
preferences over an election cycle is due to sampling error. Second, as we saw above, we lack
anything approaching a regular, daily time series of party preference until the most recent
elections, and even then only towards the end of the cycle. One cannot pursue standard
time-series tests using series of irregular and unreliable observations.
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)76 C H R I S T O P H E R W L E Z I E N ET AL.
Our solution is to analyse the data not as a set of time series but as a set of cross-sections
(following Wlezien and Erikson, 2002).The units of analysis are party preferences (observed
and imputed) for each of the last 750 days of the seventeen elections. Instead of seventeen
haphazard time series, we analyse 750 cross-sections, where we predict the actual election
result from the polls on dates t = 1 to t = 750. For each of the many days without polls
centred on those dates, we generate values through linear interpolation between the values
of the nearest available dates.13 Given that our methodology is explicitly cross-sectional, our
interpolation does not pose any of the problems that would accompany time-series analysis
of such data; indeed, interpolation offers more fine-grained evidence. With the daily
readings, we can observe: (1) the slopes (and intercepts) from these regressions; and (2) the
R-squared statistic for the model predicting the election day vote from party preferences,
which tells us the proportion of the variance in the vote explained by reported preferences
on each day.
For the cross-sectional equation predicting the actual vote from the polls on each day, we
estimate 750 equations of the following form for each party:
ELECTION DAY VOTE j = at + bt VOTE INTENTION jt + ν jt , (2)
over all j = election years, holding t (the date) fixed for each regression. We are interested
in the coefficients a and b and the prediction error as we vary t.They describe the general
relationship between the polls and the vote. It may be that these regressions suggest no
systematic change in the relationship, implying that polls evolve as a random walk over time.
Under these circumstances, we would expect a to be 0 and b to be 1 for all values of t.This
means that the polls provide the best guess of the election result at each point in time.
However, there would still be error in the predictions, presumably more so the earlier vote
intention is estimated.As the cycle unfolds and events cause preferences to change, the polls
would increasingly converge on the final result.Thus, while the coefficient (b) for the poll
variable in Equation 2 would be a constant 1.0, the explained variance (adjusted R-squared)
of the equation would increase over the election cycle.
There may not be an identity relationship (a = 0 and b = 1) between the vote intention
estimates and election results, however. If there is regression to the mean, a would be greater
than 0 and b would be less than 1. With b less than 1 and with a and b held constant, the
equilibrium is equal to a/(1-b). Thus, if the polls were greater (lesser) than a/(1-b), we
would expect the election day vote to be lower (higher) than the current polls. Assuming
the coefficients remain fairly constant over time, the explained variance will remain fairly
constant until late in the cycle, when the effects of events do not fully decay before election
day – as they increasingly affect the outcome, the R-squared increases towards the end.14 Of
course, as noted above, the equilibrium could change over time. If this were the case, we
would observe a growing adjusted R-squared over time, as a portion of the new shocks to
preferences persist to affect the outcome, with an added surge from late-arriving short-term
shocks at the end (for more details, see Erikson and Wlezien, 2012;Wlezien and Erikson,
2002; see also Campbell, 2000). Under these circumstances, we would also expect b to
increase (and a to decrease) over the cycle, as later polls become a better predictor of the
outcome.
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)P O L L S A N D T H E VO T E I N B R I T A I N 77
The Polls and the Vote in Britain
Consider the scatter plots between poll results at various points of the election cycle and the
actual party vote.15 These are shown in Figures 3–5. In the upper left-hand panel of the
figure, using the handful of elections for which polls are available 1,500 days before election
day, we see that there is little pattern.As we turn to more current polls, moving horizontally
and then vertically through the table, pattern emerges; simply, the later the polls the greater
the focus. As we get closer to the election, the polls tell us more about the outcome.This
is as one would expect.
R-squareds
In Figure 6, we make a more fine-grained presentation. Here we display the cross-sectional
R-squareds (across elections) from regressing each party’s share of the vote on their share
in the polls for each date starting 750 days before the election, when we have polls for 14
of the 17 elections.16 Specifically, for each of the three main parties we estimate the
following equation:
ELECTION DAY VOTE j = at + bV
t jt + v jt , (3)
where V is the reported vote intention in the polls at time t and each regression is estimated
over all j (election years) for fixed t (the date). Recall that we are interested in the pattern
of R-squareds and regression coefficients as we vary t.
The series of R-squareds in Figure 6 corroborate what we saw in the scatter plots. First,
early British polls contain a good amount of information about the vote, as we hypoth-
esised. Polls from two years in advance of election day account for over 33 per cent (on
average across the three parties) of the variance in the vote, suggesting that elements of the
British political system help to structure choice well in advance of the formal electoral
cycle. Second, the predictability of the vote from the polls increases during the election
cycle. Polls from one year in advance account for 50 per cent of the party vote variance, on
average, and this increases to about 90 per cent using polls from just before election day.17
These findings are consistent with the first hypothesis that polling in Britain’s parliamentary
system should be informative early in the cycle.
There are differences across the political parties, however. Most notably, the polls come
into focus more quickly for the Conservatives than for the other two parties, particularly
Labour. Concentrating on the last year in the cycle, when differences across the parties are
evident, we can see that the increase in predictability for the Conservatives is essentially
linear through to election day. For the other parties, the improvement comes towards the
end of the cycle. Notably, the pattern is not significantly different for Labour and the
Liberals during the last 100 days, with the polls becoming increasingly informative at a
similar rate near to the election.18 At the very end, the Liberal poll share predicts its party’s
vote even better than the Conservative poll standing predicts its party vote. Labour support
is a less reliable predictor through most of the final year, and especially in the months leading
up to the election, but exhibits a similar dynamic to what we observe for the Liberals –
electoral preferences are uncertain until late in the cycle but rapidly come into focus during
the final few months.
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)Figure 3: Scatter Plot of Conservative Party Vote Share by Poll Share for Selected Days of the Electoral Calendar, 1950–2010 78
1500 Days 1200 Days 900 Days 750 Days
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
30 30 30 30
20 20 20 20
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
600 Days 450 Days 350 Days 250 Days
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
30 30 30 30
20 20 20 20
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
150 Days 100 Days 75 Days 50 Days
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
30 30 30 30
20 20 20 20
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
30 Days 20 Days 10 Days 1 Day
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
30 30 30 30
20 20 20 20
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
POLITICAL STUDIES: 2013, 61(S1)
C H R I S T O P H E R W L E Z I E N ET AL.
© 2013 The Authors. Political Studies © 2013 Political Studies AssociationFigure 4: Scatter Plot of Labour Party Vote Share by Poll Share for Selected Days of the Electoral Calendar, 1950–2010
1500 Days 1200 Days 900 Days 750 Days
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
30 30 30 30
20 20 20 20
POLITICAL STUDIES: 2013, 61(S1)
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
600 Days 450 Days 350 Days 250 Days
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
© 2013 The Authors. Political Studies © 2013 Political Studies Association
30 30 30 30
20 20 20 20
P O L L S A N D T H E VO T E I N B R I T A I N
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
150 Days 100 Days 75 Days 50 Days
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
30 30 30 30
20 20 20 20
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
30 Days 20 Days 10 Days 1 Day
60 60 60 60
50 50 50 50
40 40 40 40
Vote
Vote
Vote
Vote
30 30 30 30
20 20 20 20
20 30 40 50 60 20 30 40 50 60 20 30 40 50 60 20 30 40 50 60
Polls Polls Polls Polls
79Figure 5: Scatter Plot of Liberal Party Vote Share by Poll Share for Selected Days of the Electoral Calendar, 1950–2010 80
1500 Days 1200 Days 900 Days 750 Days
40 40 40 40
30 30 30 30
20 20 20 20
Vote
Vote
Vote
Vote
10 10 10 10
0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Polls Polls Polls Polls
600 Days 450 Days 350 Days 250 Days
40 40 40 40
30 30 30 30
20 20 20 20
Vote
Vote
Vote
Vote
10 10 10 10
0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Polls Polls Polls Polls
150 Days 100 Days 75 Days 50 Days
40 40 40 40
30 30 30 30
20 20 20 20
Vote
Vote
Vote
Vote
10 10 10 10
0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Polls Polls Polls Polls
30 Days 20 Days 10 Days 1 Day
40 40 40 40
30 30 30 30
20 20 20 20
Vote
Vote
Vote
Vote
10 10 10 10
0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Polls Polls Polls Polls
POLITICAL STUDIES: 2013, 61(S1)
C H R I S T O P H E R W L E Z I E N ET AL.
© 2013 The Authors. Political Studies © 2013 Political Studies AssociationP O L L S A N D T H E VO T E I N B R I T A I N 81
Figure 6: Adjusted R-Squared from Predicting each Party’s Vote Share from its Poll Share,
by Date, the Final 750 Days
ConservaƟve Labour
1 1
Adjusted R-Squared
Adjusted R-Squared
.8 .8
.6 .6
.4 .4
.2 .2
0 0
750 600 450 300 150 0 750 600 450 300 150 0
Days UnƟl ElecƟon Days UnƟl ElecƟon
Liberals
1
Adjusted R-Squared
.8
.6
.4
.2
0
750 600 450 300 150 0
Days UnƟl ElecƟon
Regression Coefficients
The regression coefficients in Figures 7 and 8 reveal a slightly different pattern to what we
saw in Figure 6. Well before election day – 500 or more days in advance – the slope
coefficients in Figure 7 are almost constant at about 0.5 on average. As we can also see in
Figure 8, the intercepts from the regressions during this period are all positive. This set of
results implies an equilibrium vote share for each party towards which its support tends to
gravitate.The equilibrium shares look to be fairly constant during this period.We know this
because the estimated a and b coefficients do not fluctuate from 500 to 750 days before
election day. In Figure 6 we can see that poll preferences during the period also do not
improve the predictability of the vote, that is, the R-squareds do not increase.
Over the final year and a half of the cycle, the slope coefficients increase, though
differently for the three parties. The growth in coefficients tells us that polls increasingly
match the final vote for all parties. Even at the end of the cycle, however, there is an identity
relationship – that is, a one-to-one match – between the polls and the vote only for the
Liberals. With the slope coefficient less than 1.0 for the Conservative and Labour Parties,
there is regression to the mean in the final pre-election polls.19 When the party’s poll share
is above the equilibrium value, the party tends to underperform the polls on election day;
when the poll share is below that value, the party tends to do better than the polls suggest.20
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)82 C H R I S T O P H E R W L E Z I E N ET AL.
Figure 7: Regression Coefficient Predicting each Party’s Vote Share from its Poll Share, by
Date, the Final 750 Days
ConservaƟve Labour
1 1
.8 .8
.6 .6
Beta
Beta
.4 .4
.2 .2
0 0
750 600 450 300 150 0 750 600 450 300 150 0
Days UnƟl ElecƟon Days UnƟl ElecƟon
Liberals
1.5
1.2
.9
Beta
.6
.3
0
750 600 450 300 150 0
Days UnƟl ElecƟon
This may provide evidence of a systematic ‘shy voter’ effect whereby some supporters of
the less popular of the two big parties claim to be undecided, or supporting another party,
in polls but then swing behind the unpopular party on election day. Such a hypothesis was
popular after the surprisingly strong performance of the Conservatives in 1992, and some
evidence for it can be seen in subsequent elections. Election eve polls consistently overstated
Labour’s support during the period from 1997 to 2005 when the party led the Conser-
vatives, but in 2010, when Labour trailed in the polls, all of the election eve polls
underestimated Labour support (Kavanagh and Cowley, 2010, p. 252).21 There are other
possible explanations for the patterns in Figures 7 and 8. For example, it may be that
undecided voters split towards the underdog(s) later in the cycle.There are also more prosaic
methodological explanations.22
In sum, our analyses of the polls and the vote in Britain reveal a number of things. First,
we learn that polls from early in the election cycle contain a lot of information about the
election day vote. Indeed, polls from one year in advance account for, on average, half of the
variance in results across the seventeen elections between 1950 and 2010. Second, we also
learn that polls become more informative as we approach the election. Party preferences
are, in fact, largely formed by the time the official campaign begins. Preferences do still
change during the campaign and these changes can impact the actual vote, however.Third,
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)P O L L S A N D T H E VO T E I N B R I T A I N 83
Figure 8: Intercept Predicting each Party’s Vote Share from its Poll Share, by Date, the
Final 750 Days
ConservaƟve Labour
40 40
30 30
Intercept
Intercept
20 20
10 10
0 0
-10 -10
750 600 450 300 150 0 750 600 450 300 150 0
Days UnƟl ElecƟon Days UnƟl ElecƟon
Liberals
40
30
Intercept
20
10
0
-10
750 600 450 300 150 0
Days UnƟl ElecƟon
we see that the pattern differs across parties during the year leading up to the election.
Conservative Party preferences tend to evolve fairly consistently over the period whereas
Labour and Liberal preferences come into focus late in the cycle, over the final three to four
months.The results fit with our expectations in Hypothesis 1 but the differences between
parties are not those we anticipated in our motivation for Hypothesis 2 – during the official
campaign, we do not see a more dramatic increase in the informativeness of polls for the
Liberals than both of the major parties.
Comparing Britain and the US
Earlier in the article, we hypothesised that the relationship between the polls and the vote
will differ across different political systems. Specifically, we posited that preferences in the
British parliamentary system will come into focus more quickly by comparison with those
in presidential systems.23 To begin to address this possibility, we compare patterns in Britain
with those for presidential and congressional elections in the US. For this comparison,
Figure 9 shows the mean adjusted R-squared for the three main British political parties
from Figure 6. It also shows the adjusted R-squareds predicting the national presidential and
congressional vote from national polls in elections since 1948 – specifically, we predict the
Democratic share of the two-party vote from its two-party poll share.The presidential ‘trial
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)84 C H R I S T O P H E R W L E Z I E N ET AL.
Figure 9: The Predictability of Elections from Polls: Britain and the US Compared
heat’ vote question asks survey respondents which candidate they would vote for if the
election were held today.24 The congressional vote question asks which party’s candidate
for Congress in their district they would vote for, which is often referred to as the ‘generic’
ballot.25 We only report numbers going as far back before an election as is reasonable given
the availability of polls – 450 days in advance for congressional elections and 300 days back
for presidential elections.26
In Figure 9 it is clear that at the end of the election cycle, specifically the last 90 days, polls
predict the presidential vote better than the British parliamentary vote, but further away
from the election their performance drops sharply. Indeed, at the beginning of the election
year, presidential polls contain almost no information about the election day vote. As we
hypothesised (H3), parliamentary polls provide more information about election outcomes
early in the cycle and this difference disappears (quickly) once both presidential candidates
are known.The figure also supports our hypothesis that polls in centralised, parliamentary
systems such as Great Britain will be substantially more informative than legislative polls in
more candidate-oriented presidential systems such as the US: British polls are substantially
more informative than congressional polls.We cannot be sure of the precise mechanism, but
factors such as the greater role played by incumbency or candidate quality in the more
localised US elections and the role that assessments of national parties and their leaders play
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)P O L L S A N D T H E VO T E I N B R I T A I N 85 in structuring public opinion in the more centralised Westminster elections may help explain the difference. Discussion and Conclusion Our examination of polls and the vote in Britain reveals a substantial pattern that comports with our conjectures. To begin with, we confirmed our expectation that a significant amount of information about parties’ prospects in Britain’s parliamentary system is con- tained in polls conducted even a year or more before election day.We also confirmed party differences in the evolution of electoral preferences over the course of the election cycle. We find that Conservative electoral performance starts to come into focus earlier, and this increases continuously from eighteen months out. By contrast, the Labour and Liberal vote only begins to sharpen in the polls around five months before election day, and even then expressed preferences are less informative than for the Conservatives. Further substantive work is needed to understand the reasons for these patterns and explore explanations for the other differences we observe across parties. There are other questions as well. It might be that preferences evolve differently for governing and opposition parties – such as those suggested by theories of retrospective voting (Fiorina, 1981). The comparison of Britain and the US confirmed our expectation that British polls are much more informative at early stages of the electoral cycle than American polling: in congressional races this holds throughout the cycle and in presidential elections it is true until the last months of the cycle, when both presidential candidates are known. The nationalised, party-centred campaigns and stable party leadership found in Britain are reflected in more stable preferences, which reveal an underlying structure to vote intention polls very early in the electoral cycle. British electoral preferences are not perfectly stable, however. They change over the course of a cycle and a substantial portion of these changes lasts to impact the outcome. The evolution begins about one and a half years before election day, well in advance of the formal election campaign. It continues up to election day.We cannot pinpoint what actually drives this evolution.What we can say is that something happens over this eighteen-month ‘long campaign’ to crystallise preferences gradually. Our investigation also points towards questions for future research in wider comparative context. Of obvious importance is whether the differences we identify between the US and Britain are generalisable across other presidential and parliamentary countries.We also want to assess the differences within these government types.The usual suspects are electoral and party systems, but other aspects of context, such as federalism, might also matter (see Wlezien, 2010). This requires additional theorising and analysis. For now, the contrast between Britain and the US is strongly suggestive of substantive differences in the evolution of electoral preferences across political systems. Appendix The data used for this article consist of 3,293 opinion polls conducted in Britain between 1945 and 2010 by eighteen different polling houses: Gallup, National Opinion Polls (NOP), Business Decisions, Opinion Research Centre (ORC), Research Services Ltd (RSL), ASL, Market & Opinion Research International (MORI)/Ipsos-MORI, Marplan/ICM © 2013 The Authors. Political Studies © 2013 Political Studies Association POLITICAL STUDIES: 2013, 61(S1)
86 C H R I S T O P H E R W L E Z I E N ET AL.
Research Ltd, Harris, Harris Interactive, Communicate Research (ComRes), Populus,
Rasmussen,YouGov, BPIX, Nielsen Media Research (NMR), Angus Reid and Opinium.
These data were compiled from a large number of sources, with further cross-checks and
triangulation in the case of inconsistencies or missing data (in particular with reference to
missing sample sizes). Further details of the collection of poll data are reported below.
Much of the poll data for the 1997–2001, 2001–5 and 2005–10 election cycles are
compiled from Ipsos-MORI’s (2010b) ‘Voting intentions (Westminster) – All Companies’
Polls’ series which reports all polls published on GB vote intention by major pollsters
and/or from the cross-tabs published online by polling houses. For the election cycles prior
to 1997 poll data were compiled from, and cross-checked against, a number of different data
series.
• Gallup: data on 811 polls conducted by the Gallup Organisation between 1950 and 2000
were compiled from the original Gallup Political and Economic Index and from King and
Wybrow’s (2001) British Political Opinion 1937–2000, and supplemented with poll data
provided by Michael Thrasher. Additional data were also compiled from National
Opinion Polls’ Political Bulletin and its Political, Social, Economic Review. Further cross-
checks were made against the fieldwork dates of reports in the Roper Centre for Public
Opinion Research. Gallup ceased political polling in Britain in 2001.
• National Opinion Polls: data on 530 polls conducted by National Opinion Polls (NOP)
between 1959 and 2005 were compiled from data provided by Nick Moon and Michael
Thrasher and from NOP’s Political Bulletin and Political, Social, Economic Review, in
addition to data from the annual ‘Reference Section’ of the British Elections and Parties
Yearbook (formerly the British Elections & Parties Review).
• Marplan/ICM Research Ltd: data on 510 polls conducted by Marplan/ICM Research Ltd
were compiled from ICM’s (2010) ‘Guardian Voting Series’, published online, and from
data provided by Michael Thrasher.
• Market & Opinion Research International/Ipsos-MORI (from 2005): data on 673 polls
conducted by MORI (now Ipsos-MORI) between 1976 and 2010 are compiled from
Ipsos-MORI’s (2010a) Voting Intention in Great Britain, published online.
• Harris: data on 141 polls conducted by Lou Harris & Associates were compiled from data
provided by Michael Thrasher. Note that Harris Interactive is treated as a distinct polling
house.
Where possible, poll data were cross-checked and triangulated against other available
sources, in particular the annual ‘Reference Section’ of the British Elections and Parties
Yearbook (formerly the British Elections & Parties Review), over the period between 1987 and
2005 by David Denver, David Broughton, Justin Fisher, Colin Rallings, Michael Thrasher,
Phil Cowley, Andrew Russell and others. We still lack historic poll data from the 1950s,
1960s and 1970s for Opinion Research Centre (ORC), Lou Harris and Business Decisions
(pollster for The Observer in the 1970s), Research Services Limited (RSL) and for ASL
during the 1990s, while exact fieldwork dates are often not available for some NOP polls
from the 1950s, 1960s and 1970s. Tracking down these data would be desirable, but at
present it is unclear whether records still exist other than in newspaper reports – in
particular of fieldwork dates and sample sizes.
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)P O L L S A N D T H E VO T E I N B R I T A I N 87
Table A1: Number of Polls by Polling House, 1950–2010
Polling house Polls
1 Gallup 811
2 National Opinion Polls (NOP) 530
3 MORI/Ipsos-MORI 673
4 Marplan/ICM Research Ltd 510
5 Harris 141
6 Harris Interactive 16
7 Communicate Research (ComRes) 87
8 Populus 94
9 Rasmussen 3
10 YouGov 279
11 BPIX 33
12 Nielsen Media Research (NMR) 10
13 Angus Reid 20
14 Opinium 10
15 Business Decisions 4
16 Opinion Research Centre (ORC) 8
17 Research Services Limited (RSL) 4
18 ASL 59
3,293
Poll Questions
• Gallup: ‘If there were a general election tomorrow, which party would you support?’,
‘Which party are you most inclined to vote for?’
• NOP: ‘If you do vote in the next general election, which party will you vote for?’
• Ipsos-MORI: ‘How would you vote if there were a General Election tomorrow?’
• ICM Research Ltd: ‘If there were a general election tomorrow which party do you
think you would vote for?’
• Harris Interactive:‘If the general election were held tomorrow for which party do you
intend to vote?’
• Communicate Research (ComRes): ‘If there were a general election tomorrow, would
you vote Conservative, Labour, Liberal Democrat or some other party?’
• Populus: ‘If the general election was tomorrow, which party would you vote for?’
• YouGov:‘If there were a general election tomorrow, which party would you vote for?’
• Angus Reid: ‘If a General Election were held tomorrow, which one of the following
parties would you be most likely to support in your constituency? – ‘Decided Voters
with Leaners’ (March 2010), ‘In the General Election that will take place on 6 May,
which one of the following parties are you most likely to support in your constituency?’
(April 2010).
• Opinium:‘If there were a general election tomorrow, for which party would you vote?’
(Accepted: 30 July 2012)
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)88 C H R I S T O P H E R W L E Z I E N ET AL.
About the Authors
Christopher Wlezien is Professor of Political Science at Temple University. He has published widely in the areas of
elections, public opinion and public policy, and his books include Degrees of Democracy (2010), Who Gets Represented?
(2010), and The Timeline of Presidential Elections (2012). He was founding editor of the Journal of Elections, Public Opinion
and Parties and is co-editor of ‘Poll-Reviews’ for Public Opinion Quarterly. Christopher Wlezien, Department of Political
Science, Temple University, Philadephia, PA 19122-6089, USA; email: wlezien@temple.edu
Will Jennings is Senior Lecturer in Politics at the University of Southampton. He specialises in the quantitative
analysis of politics, policy and society (in particular in the application of time-series methods) as well as in the field
of executive politics and the governance of risk in mega-projects. Will Jennings, Department of Politics and
International Relations, School of Social Sciences, University of Southampton, Southampton SO17 1BJ, UK; email:
W.J.Jennings@soton.ac.uk
Stephen Fisher is the University Lecturer in Political Sociology at the University of Oxford, and the Fellow and Tutor
in Politics at Trinity College. He is the author of various articles on elections, voting, opinion polls and social attitudes.
Stephen Fisher, Trinity College, University of Oxford, Oxford OX1 3BH, UK; email: stephen.fisher@trinity.ox.ac.uk
Robert Ford is Lecturer in Politics at the University of Manchester. His research covers public opinion, race and racial
attitudes, radical right political parties and voting analysis. He has published in many international journals including
Electoral Studies, the European Journal of Political Research and the British Journal of Sociology. He is currently working on
a forthcoming book on the radical right in Britain, co-authored with Matthew Goodwin (University of Nottingham).
Robert Ford, Department of Politics, Arthur Lewis Building, University of Manchester, Manchester M13 9PL, UK;
email: rob.ford@manchester.ac.uk
Mark Pickup is an Assistant Professor in the Department of Political Science at Simon Fraser University. He is a
specialist in comparative politics, with a particular interest in public opinion and democratic representation. His
research focuses on political information, public opinion, the media, election campaigns and electoral institutions.
Mark also has methodological interests in the analysis of longitudinal data. He teaches introductory and advanced
time-series courses at the ICPSR summer programme. His work has appeared in several venues, including the Journal
of Politics, the Journal of Conflict Resolution, Electoral Studies, the Canadian Journal of Political Science and the British Journal
of Political Science. Mark Pickup, Department of Political Science, Simon Fraser University, 8888 University Drive,
Burnaby BC V5A 1S6, Canada; email: mark.pickup@gmail.com
Notes
An earlier version of this article was presented at the Annual Meeting of the Elections, Public Opinion and Parties (EPOP) subgroup
of the Political Studies Association, University of Essex, Colchester, 10–12 September 2010. We thank Michael Thrasher, David
Cowling, Nick Moon and Mark Pack for their help with various poll data.
1 It is important to note that the late crystallisation of support for the Liberals would, by construction, have implications for the
crystallisation of support for another party, or parties (including ‘others’), since late shifts in support must come from (or go)
somewhere.Technically voters do not have to shift support from party to party, they also can shift from ‘don’t know’ or ‘won’t say’
to a party or vice versa.
2 We only include polls where the exact fieldwork dates are known (this leads to exclusion of NOP polls from the period between
1959 and 1964 where only the month of fieldwork is known). ‘Liberal’ polls refer to the Liberal Party prior to 1981, to the
aggregated vote intention for the SDP and Liberal Party for the period between 1981 and 1990, which formed a formal electoral
pact, the ‘SDP–Liberal Alliance’ between 1981 and 1988, and to Liberal Democrat support after 1990.
3 Although these decision rules may seem innocuous enough, they have been shown to affect the tally, particularly in recent election
years (Erikson et al., 2004;Wlezien and Erikson, 2001). For example, likely voter samples sometimes distort the reported vote prior
to election day by exaggerating short-term trends in who is sufficiently ‘likely’ to vote. Still, we usually have little choice but to
rely on results from voter samples used for the headline figures as these are the only figures we can reliably obtain for most
historical polling time series.
4 For surveys in the field for an even number of days, the fractional midpoint is rounded up to the following day.
5 It is important to note that polls on successive days are not truly independent. Although they do not share the same respondents,
they do share overlapping polling periods. Thus polls on neighbouring days will capture a lot of the same things, which is of
consequence for an analysis of dynamics.
6 Previous work in different countries (Erikson and Wlezien, 1999; Fisher et al., 2011; Jackman, 2005; Pickup and Johnston, 2007)
has demonstrated that within a given electoral cycle, the polls from a given polling house may contain a constant error – a house
effect. In order to evaluate the relationship between true public opinion and the electoral outcome, it may be desirable to correct
polls for these house effects. Unfortunately, the sparseness of data on polls in the early electoral cycles makes the estimation of these
house effects problematic. Since we are unable to produce adjustments for every cycle, we choose not to adjust the polls in any
cycle. However, the literature cited above shows that house effects are small compared with industry bias, which itself has varied
dramatically between elections in the UK.
7 There are no polls during the last 30 days for the 1950 and 1951 elections.
© 2013 The Authors. Political Studies © 2013 Political Studies Association
POLITICAL STUDIES: 2013, 61(S1)You can also read
